ScoutB
Senior Member (Voting Rights)
I agree it looks AI generated for the same reasons @ME/CFS Science Blog mentioned. Also a couple of the phrases seem like the kind of "sounds smart but is actually devoid of meaning" writing AI tends to produce.
Preprint Biological Insights from Genome-Wide Association Studies and Whole Genome Sequencing of [ME/CFS], 2026, Maccallini et al
- Thread starter forestglip
- Start date
We have split off a strand of conversation to Very mild or prodromal ME/CFS.
We have split off the discussion of Paolo's latest work with the dendrogram of disease similarities to a new thread Genetic similarities between ME/CFS and other diseases
We have split off the discussion about the possibility of permanent degeneration including neuronal loss in ME/CFS to an existing thread
Does ME/CFS cause permanent damage? Is it a degenerative disease ? Discussion Thread
If anyone is concerned that the moving of posts has caused a lack of context that affects their posts, do get in touch with a moderator.
Does ME/CFS cause permanent damage? Is it a degenerative disease ? Discussion Thread
If anyone is concerned that the moving of posts has caused a lack of context that affects their posts, do get in touch with a moderator.
Last edited:
Hoopoe
Senior Member (Voting Rights)
Hi @paolo,
If eMSN are the most important cell type in the illness, then one would expect them to be involved in functions that are closely related to the symptoms of the illness. Can we draw a connection between eMSNs and the more specific symptoms of ME/CFS?
Personally I view my illness as a form of sensitivity to exertion, where exertion can mean physical activity, mental activity like socializing and concentrating, or being exposed to sensory stimuli. My ability to bear this exertion declines rapidly and some kind of intolerance builds up. This can occur in the form of next-day PEM, but also in a more subtle way over days, weeks, months. I believe that my illness involves some problem in a part of human physiology that is important for allowing a person to bear exertion, to recover from it, to adapt to it. This part also has to be still poorly understood or its importance not recognized, or we would probably already know its importance in ME/CFS.
What physiology exactly is involved is difficult to guess. It might be a problem with sleep, or something to do with synaptic adaptation to exertion, or coordination of post-exertional responses, or processing of inputs related to exertion, things like that.
What I would like to know if there is some way to figure out what exactly the eMSN are doing and whether they are involved in any of these functions.
If eMSN are the most important cell type in the illness, then one would expect them to be involved in functions that are closely related to the symptoms of the illness. Can we draw a connection between eMSNs and the more specific symptoms of ME/CFS?
Personally I view my illness as a form of sensitivity to exertion, where exertion can mean physical activity, mental activity like socializing and concentrating, or being exposed to sensory stimuli. My ability to bear this exertion declines rapidly and some kind of intolerance builds up. This can occur in the form of next-day PEM, but also in a more subtle way over days, weeks, months. I believe that my illness involves some problem in a part of human physiology that is important for allowing a person to bear exertion, to recover from it, to adapt to it. This part also has to be still poorly understood or its importance not recognized, or we would probably already know its importance in ME/CFS.
What physiology exactly is involved is difficult to guess. It might be a problem with sleep, or something to do with synaptic adaptation to exertion, or coordination of post-exertional responses, or processing of inputs related to exertion, things like that.
What I would like to know if there is some way to figure out what exactly the eMSN are doing and whether they are involved in any of these functions.
Last edited:
I don't know how eMSNs can cause the specific symptoms of ME/CFS, like PEM and also orthostatic intolerance.
I noted that the other cell type reported as a hit in my analysis has received little consideration here. White matter neurons (WMNs, also called interstitial white matter neurons, IWMNs) seem to play a role in brain circulation. Something I am thinking about is that the regulation of brain blood flow may be implicated in PEM, and I found this model in one of the reviews that I used in my paper (link) (from the second paragraph of the discussion).
I don't have a model for the disease; I think it is still premature: in the manuscript I did not propose a model. I am probably biased toward the glutamatergic system, and I try to destroy my biases with a constant effort. My hope is that the disease model will ultimately arise almost entirely from integrating experimental data from ME/CFS patients with structured databases (such as scRNA-seq brain atlases and other resources). I am working in this direction.
Last edited:
Hoopoe
Senior Member (Voting Rights)
Interesting. They are also involved in sleep. It sounds like this could be related to difficult tolerating exertion in the short term (blood flow) and longer term (sleep).
ME/CFS Science Blog
Senior Member (Voting Rights)
I was able to replicate Paolo's meta-analysis (MVP + DME_1) using MungeSumstats R package (takes a while to load!) and METAL as the paper describes.
I got the same number of SNPs (8,859,361) and top hits. Sometimes the z-sign is flipped but don't think that matters much. I noticed, however, that I get quite different results for the cell type and gene set analyses with MAGMA if I set the window to 35,10 as some authors do, rather than using 0,0 - the conservative option that Paolo used. These windows are how far away a SNPs can be in order to be counted in support of a particular gene.
The 35,10 window was, for example, used in the Duncan et al. 2025 paper. I thought I would get stronger results this way but the opposite was true. It's best visible in a graphs below.
Here's what the results look like with window 0,0 almost (see note*) exactly the same as the results of Paolo. These are the 31 cell types normalised per dissection with the top hits for eMSN.
But with window 35,10 the top hits were for upper_layer_intratelencephalic cells while oligodendrocyte_precursor jumped up as well.
I suspect this is due to FUMA's approach of normalizing the gene expression per dissection. These make for small datasets where particular genes can have a lot of influence. The Duncan et al. 2025 takes a different approach by analyzing at 461 cell types and normalizing by gene expression in the entire dataset.
When I applied this to the meta-analysis, the results seem less affected by the window size.
Window 0,0
Window 35,10
* I got the same results as Paolo if I use the FUMA website but since that takes longer and it sometimes fails, so I do the analysis locally on my own laptop. This results in tiny differences in SNP handling possibly due to different MAGMA versions used. But the results are basically the same. For the Siletti level 2 brain atlas, for example the correlation in p-values between the two methods is 0.987.
I got the same number of SNPs (8,859,361) and top hits. Sometimes the z-sign is flipped but don't think that matters much. I noticed, however, that I get quite different results for the cell type and gene set analyses with MAGMA if I set the window to 35,10 as some authors do, rather than using 0,0 - the conservative option that Paolo used. These windows are how far away a SNPs can be in order to be counted in support of a particular gene.
The 35,10 window was, for example, used in the Duncan et al. 2025 paper. I thought I would get stronger results this way but the opposite was true. It's best visible in a graphs below.
Here's what the results look like with window 0,0 almost (see note*) exactly the same as the results of Paolo. These are the 31 cell types normalised per dissection with the top hits for eMSN.
But with window 35,10 the top hits were for upper_layer_intratelencephalic cells while oligodendrocyte_precursor jumped up as well.
I suspect this is due to FUMA's approach of normalizing the gene expression per dissection. These make for small datasets where particular genes can have a lot of influence. The Duncan et al. 2025 takes a different approach by analyzing at 461 cell types and normalizing by gene expression in the entire dataset.
When I applied this to the meta-analysis, the results seem less affected by the window size.
Window 0,0
Window 35,10
* I got the same results as Paolo if I use the FUMA website but since that takes longer and it sometimes fails, so I do the analysis locally on my own laptop. This results in tiny differences in SNP handling possibly due to different MAGMA versions used. But the results are basically the same. For the Siletti level 2 brain atlas, for example the correlation in p-values between the two methods is 0.987.
I found the same thing in my initial attempts to just do MAGMA on GTEx tissues. I saw that the brain tissues were much less significant than in DecodeME and @tralfamadorian97's same analyses, and I realized the biggest difference was because I used 35,10 and tralfamadorian used a 0,0 window instead. I'm not totally sure DecodeME used 0,0 too, but I think that's the default on FUMA, which was the platform they used.
I had assumed 35,10 would be even more significant, since so many of the top hits seem to be loci upstream of genes.
hotblack
Senior Member (Voting Rights)
I have vague memories of this but do not know and am not sure I now or ever properly understood it or why it has this effect. Could you or @ME/CFS Science Blog explain any more?
ME/CFS Science Blog
Senior Member (Voting Rights)
MAGMA connects SNP associations with the disease (ME/CFS) to genes. If a couple of SNPs are located inside the gene, their p-values are combined into a gene signal. The window option lets you include SNPs into the gene signal that are not in but just next to the gene.
I thought that allowing a little bit of window would help to get a stronger signal because otherwise a lot of SNPs remains unused in MAGMA, but the opposite seems to be true. The window seems to add mostly noise.
Thank you for performing the replication. I see that using the 461 cell types (the Duncan approach) gives more robust results, even though we lose the anatomical resolution in most cases. Perhaps, at this stage, this method is better than the FUMA L2 database (the one I used).
I tend to think that DecodeME used a 0 window in MAGMA, because it is the default setting in FUMA website (which is what they used, according to the preprint). But the preprint does not explicitly indicate the window, as far as I can tell.
ME/CFS Science Blog
Senior Member (Voting Rights)
One very minor thing: could it be that you used MAGMA v.1.08 rather than v1.10? Because 1.08 is what the FUMA SNP2GENE website uses when I try it. When I used MAGMA v.1.10 on my laptop without the FUMA website, I got slightly different results.
In my pipeline, MAGMA and cell-type analysis are performed by the FUMA platform. In other words, I did not use a local installation. I used the following versions: FUMA = v1.8.3 and MAGMA = v1.10. This is specified in paragraph 2.5 of the preprint. Also, I included the params.config file of each FUMA job in a folder of the associated GitHub repository.
ME/CFS Science Blog
Senior Member (Voting Rights)
Background
Had a closer look at the conditional analysis as specified by Watanabe et al. 2019. Step 3 is about checking if cell types from different datasets are independent of each other.
The basic idea is that for each pair of cell types, you first do a MAGMA regression with one cell type, but with the averages from both datasets added. This gives the marginal p-value for the two cell types in the pair. Then you do the same two regressions but add in the other cell type as a covariate, resulting in a conditional p-value for the two cell types. The proportional significance (PS) indicates how much the p-value became less significant when you condition on the other cell type: PS = -log10(p_conditional) / -log10(p_marginal).
With that info, Watanabe uses a simple rule to test if the cell types are independent: start with the most significant cell type and check if others' PS stays above 0.5 (then move to the second cell type, etc.).
Analysis
Now, on to the Maccalini paper. The most significant cell type (the only one that survived Bonferroni correction) was the eMSN from: 47_Siletti_CerebralNuclei.Cla_Human_2022_level2
If I check the 3/19 Seeker white matter cells and test them with the eMSN from the 47 dataset as a covariate, their PS values were < 0.5, meaning they were not really independent.
In my attempt at implementing the Watanabe algorithm, I got three independent cell types:
But the last two their PS when conditioned on the top cell type, were only just above 0.5 (0.575 and 0.501).
So, in short, not much evidence of independent cell types or signals, unfortunately.
Had a closer look at the conditional analysis as specified by Watanabe et al. 2019. Step 3 is about checking if cell types from different datasets are independent of each other.
The basic idea is that for each pair of cell types, you first do a MAGMA regression with one cell type, but with the averages from both datasets added. This gives the marginal p-value for the two cell types in the pair. Then you do the same two regressions but add in the other cell type as a covariate, resulting in a conditional p-value for the two cell types. The proportional significance (PS) indicates how much the p-value became less significant when you condition on the other cell type: PS = -log10(p_conditional) / -log10(p_marginal).
With that info, Watanabe uses a simple rule to test if the cell types are independent: start with the most significant cell type and check if others' PS stays above 0.5 (then move to the second cell type, etc.).
Source: Genetic mapping of cell type specificity for complex traits | Nature Communications
Analysis
Now, on to the Maccalini paper. The most significant cell type (the only one that survived Bonferroni correction) was the eMSN from: 47_Siletti_CerebralNuclei.Cla_Human_2022_level2
If I check the 3/19 Seeker white matter cells and test them with the eMSN from the 47 dataset as a covariate, their PS values were < 0.5, meaning they were not really independent.
In my attempt at implementing the Watanabe algorithm, I got three independent cell types:
But the last two their PS when conditioned on the top cell type, were only just above 0.5 (0.575 and 0.501).
So, in short, not much evidence of independent cell types or signals, unfortunately.
Jonathan Edwards
Senior Member (Voting Rights)
The analysis here is beyond what I can follow but my thought is that for any signalling disruption in the brain there are likely to be at least four cell types whose gene expression would alter the likelihood of the problem occurring. One + and one - for one half of a loop and two more for the other half, and that does not include some likely housekeeping interneurons. The three cell types above might easily form a pathway Achilles heel.
I think the search for a cell type is worthwhile but only as a clue to some sort of operational hypothesis that might turn out to work better with another set of cells.
I think the search for a cell type is worthwhile but only as a clue to some sort of operational hypothesis that might turn out to work better with another set of cells.
The independence reported in my analysis (figure 4) is between the cell-types that survive BH correction in discovery and are replicated in HEAL2: eMSN in Cerebral Nuclei BLN.La (Siletti) and the glutammatergic neurons in cerebellar white matter (Seeker). In both comparisons we have PS>0.5.
Last edited:
ME/CFS Science Blog
Senior Member (Voting Rights)
Yes, you reported it correctly, and thanks for clarifying.
I just wanted to add that the Seeker white matter cells aren't seen as independent from the top Siletti eMSN signal if you follow the Watanabe 2019 approach (EDIT: ignoring the HEAL2 replication).